skills/swarm/SKILL.md
Large-scale parallel implementation with multi-agent swarm. Architect + QA + Implementer(s) collaborate via peer DM.
npx skillsauth add thkt/claude-config swarmInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
Multi-agent swarm for large-scale implementation. Use /code for tasks under 5 files.
Use /swarm when any of these conditions apply. Otherwise use /code.
| Condition | /swarm | | --------------------- | ------ | | Files >= 5 | Yes | | Multi-domain | Yes | | Design decisions many | Yes |
Leader is the orchestrator, not a worker. Substantive work happens through peer DM between Architect, QA, and Implementer(s).
Implementation description: $ARGUMENTS (required, prompt if empty)
See ${CLAUDE_SKILL_DIR}/../_lib/sow-resolution.md
| Agent | subagent_type | Responsibility | Bash | SendMessage | Model | | -------------- | ------------------- | ----------------------------------- | ---- | ----------- | ------ | | Leader | (self) | User interface, QG, lifecycle | Yes | broadcast | opus | | Architect | architect-feature | Codebase analysis, contracts | No | peer DM | opus | | QA | team-qa | Quality observations (non-blocking) | No | peer DM | sonnet | | Implementer(s) | team-implementation | RGRC implementation | Yes | peer DM | opus |
Haiku is excluded from team agents. It cannot reliably follow multi-step instructions or handle the shutdown protocol. The Model column in Team Architecture defines each agent's model.
Peer DM transport with handoff structures (Spawn Context, Architect Output, Implementer Started/Assignment/Completion).
See: ${CLAUDE_SKILL_DIR}/references/contracts.md#context-contracts
| Phase | Action | Detail | | ----- | --------------------------- | ------------------------------------------------------------------------------ | | 0 | SOW Detection | ${CLAUDE_SKILL_DIR}/references/execution.md#phase-0-sow-detection | | 1 | Team Setup + Architecture | ${CLAUDE_SKILL_DIR}/references/execution.md#phase-1-team-setup--architecture | | 2 | Decomposition Approval | ${CLAUDE_SKILL_DIR}/references/execution.md#phase-2-decomposition-approval | | 3 | Test Generation | ${CLAUDE_SKILL_DIR}/references/execution.md#phase-3-test-generation | | 4 | File Assignment | ${CLAUDE_SKILL_DIR}/references/execution.md#phase-4-file-assignment | | 5 | RGRC Implementation | ${CLAUDE_SKILL_DIR}/references/execution.md#phase-5-rgrc-implementation | | 6 | Integration + Quality Gates | ${CLAUDE_SKILL_DIR}/references/execution.md#phase-6-integration--quality-gates | | 7 | Summary | ${CLAUDE_SKILL_DIR}/references/execution.md#phase-7-summary |
Phase 5 must spawn all Implementers as concurrent Task calls within a single response. One Task call per unit. Sequential spawning defeats the swarm model and stretches wall time linearly.
| Scenario | Action |
| ---------------------------- | ------------------------------------------------------------- |
| User cancels | shutdown_request to all agents, TeamDelete |
| Contract fundamentally wrong | Shutdown Implementers, Architect redesign |
| Implementer no started DM | 120s timeout, shutdown, re-spawn (max 1 retry, escalate) |
| Implementer death mid-work | Leader checks worktree via git status, re-spawn (max 1 retry) |
| QG fails 3 times | Escalate to user with details |
| Agent Bash permission block | Use mode: "dontAsk" for worktree-isolated Implementers |
| test-gen timeout | Leader generates tests directly |
| test-gen produces 0 tests | Verify spec exists, ask user |
| Shutdown unresponsive | Retry with explicit tool params, move team dir to ~/.Trash/ |
Leader reports progress table at key events (Phase 4 start, Implementer started/completion, merge/QG results).
See: ${CLAUDE_SKILL_DIR}/references/contracts.md#progress-tracking
Recovery procedures by phase.
See: ${CLAUDE_SKILL_DIR}/references/contracts.md#abort--rollback
tools
Delegate implementation to codex (coder) via the herdr-agentchat plugin and drive a two-pane conversation to completion.
development
Extract recurring patterns from past closed PRs/issues and the research findings in workspace/research/, verify them against the latest code, and propose them to docs/wiki/ via PR.
development
Create Decision Records (DR) in MADR v4 format with auto-numbering.
development
Codex review + cleanup. Findings are challenged by critic-audit, not aggregated as facts, and fixes are applied directly. Do NOT use for internal multi-reviewer deep audits or findings reports (use /audit).